Quantitative analysis for identifying molecular subtypes of small cell lung cancer via two-dimensional and three-dimensional contrast-enhanced computed tomography images: a preliminary study

Introduction

Lung cancer remains a pervasive global health challenge, accounting for a substantial portion of cancer-related morbidity and mortality worldwide (1,2). Among its subtypes, small cell lung cancer (SCLC) comprises roughly 15% of lung cancer cases (3) and is characterized by aggressive behavior and poor prognosis as a high-grade neuroendocrine (NE) carcinoma (4,5). Recent advances in molecular oncology have revealed that SCLC is not a uniform entity but contains distinct molecular subtypes. Specifically, SCLC can be divided into NE subtypes—driven by the transcription factors ASCL1 (SCLC-A) and NEUROD1 (SCLC-N)—and non-NE subtypes defined by YAP1 (SCLC-Y) and POU2F3 (SCLC-P) expression (4,6). These molecular subtypes exhibit unique genetic and transcriptomic profiles that can guide treatment decisions, prognostication, and the development of targeted therapies (7-9). Notably, tumors of the NE subtype are associated with especially poor clinical outcomes (4,10). Thus, determining the molecular subtype of SCLC for each patient is of significant clinical importance.

In current practice, molecular subtyping of SCLC relies on invasive tissue biopsy followed by immunohistochemical or genetic assays, which may be infeasible in some cases. There is a clear need for noninvasive methods to infer tumor subtype. Medical imaging offers promise in this regard: it is routine, rapid, and provides whole-tumor assessment. Indeed, imaging phenotypes can reflect underlying tumor biology such as gene expression and microenvironment characteristics (11,12). In particular, radiomics—the high-throughput extraction of quantitative features from radiographic images—has emerged as a powerful approach to capture tumor heterogeneity (13). Radiomic features can quantify subtle tumor textural patterns that are imperceptible to human observers, and these features have shown good reproducibility (14). Moreover, radiomics has demonstrated value in oncology for prognostication and predictive modeling (15-17). For instance, prior studies in lung adenocarcinoma have shown that radiomics patterns derived from computed tomography (CT) can predict specific molecular alterations such as epidermal growth factor receptor (EGFR) mutations or anaplastic lymphoma kinase (ALK) rearrangements, suggesting that radiomics is capable of capturing intratumoral genetic information (18-20). However, to date, there is a paucity of research applying radiomics to SCLC, and no prior studies have addressed radiomics classification of SCLC molecular subtypes.

In this preliminary study, we investigated whether quantitative radiomics features extracted from contrast-enhanced CT (CECT) scans of primary SCLC tumors can noninvasively distinguish between the NE and non-NE molecular subtypes of SCLC. Moreover, we compared a two-dimensional (2D) radiomics analysis of a single representative tumor slice with a three-dimensional (3D) volumetric analysis to determine if the simpler 2D method can achieve similar predictive accuracy and clinical relevance. We present this article in accordance with the STARD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1041/rc).

Methods

Patient selection

This retrospective study was conducted at a single center. The study was approved by Ethics Committee of the Cancer Hospital, Chinese Academy of Medical Sciences (approval No. NCC2016G-029). Informed consent was waived due to the retrospective nature. All procedures were conducted in accordance with the Declaration of Helsinki and its subsequent amendments. We identified a total of 343 patients diagnosed with SCLC who underwent surgical resection and had available molecular subtyping results at our hospital between January 2005 and April 2016. The exclusion criteria were as follows: (I) patients who received neoadjuvant therapy (radiation, chemotherapy, etc.) before surgery; (II) patients with missing CT images; (III) patients with CT scans without contrast enhancement; (IV) patients with scans performed more than two weeks before surgery; and (V) incomplete clinical and inaccessible medical record information. The non-NE subtype (non-NE) group served as the negative group, whereas the NE group served as the positive group. This retrospective observational study was approved by our institutional ethics committee. Ultimately, a total of 51 patients meeting the inclusion criteria were included in the analysis (Figure 1).

Figure 1 Patient selection flow diagram. CT, computed tomography; SCLC, small-cell lung cancer.

Follow-up

The follow-up process encompasses all patients, commencing from the first day postoperatively. Follow-up includes regular physical examinations, chest X-ray imaging every three months, chest CT scans every six months in the first two years postoperatively, followed by chest X-ray examinations every six months, and annual chest CT scans. Occurrences of tumor recurrence were obtained by reviewing medical records or through telephone interviews by trained personnel. If unable to contact patients or their relatives, survival information from the last follow-up date was recorded as the final record. Follow-up was concluded by December 2023.

Instruments and general information

Chest CECT examinations were performed via multidetector CT scanners, including an Optima CT660, Discovery CT750 (GE Medical System, Milwaukee, WI), Revolution CT, BrightSpeed CT, and Toshiba Aquilion 64-slice spiral CT. Prior to the examination, each patient underwent breath-hold training. The following scanning parameters were applied uniformly: slice thickness/slice increment of either 1 or 1.25 mm, tube voltage of 120 kVp, and tube current ranging from 200 to 350 mAs. For contrast enhancement, 80–90 mL of iopromide (iodine concentration 300 mg I/mL) was injected intravenously at a flow rate from 2.5 to 3 mL/s, and imaging was performed 25–30 seconds after injection (arterial phase). Image reconstruction employed both standard algorithms and high-resolution algorithms, along with parallel multiplane reconstruction techniques. For image observation, the lung window (with a window width of 1,600 HU and a window level of −400 HU) and the mediastinal window (with a window width of 400 HU and a window level of 40 HU) were selected.

Conventional CT feature analysis

Two radiologists, Reader 1 and Reader 2, each with 3 years of experience in chest imaging diagnosis, collaboratively interpreted the following CT features: maximum tumor diameter; lobulation; spiculated; pleural retraction; pleural attachment; location; position; lobe; and shape. The definitions of the above semantic terms are provided in Appendix 1. In addition, patients were assessed for the presence of emphysema and interstitial pneumonia. In cases of disagreement between the two radiologists, a third reviewer, a professor in radiology with 20 years of clinical experience in thoracic imaging, made the final decision.

Image acquisition and segmentation

Digital imaging and communications in medicine (DICOM) images were retrieved from the picture archiving and communication system (PACS) and preprocessed by being resampled to a uniform resolution, thereby standardizing the image thickness and eliminating variations among images with different slice thicknesses. Z-score normalization was applied to the CECT images of each patient to achieve a consistent distribution of image intensities (21). Segmentation of regions of interest (ROI) was performed via ITK-SANP software (version 3.8.0). Lesion segmentation was performed by a dedicated imaging graduate student with expertise in thoracic imaging diagnoses. The semiautomatic method was employed for segmentation. For lesions with unclear boundaries, manual adjustments were made to refine the ROI delineated after semiautomatic outlining. The accuracy of the segmentation outcomes was verified by a senior physician specializing in thoracic imaging diagnoses. In cases of discrepancies, consensus was reached through consultation, with a deliberate effort to prevent any annotations from extending beyond the lesion boundaries.

Inter-observer and intra-observer consistency assessment

To ensure the reproducibility of ROI segmentation, both inter-observer and intra-observer consistencies were evaluated (21). A random subset of 30 patients was selected. For intra-observer assessment, the same radiologist repeated the segmentation after a washout period of at least two weeks using the identical semiautomatic delineation workflow. For inter-observer assessment, another radiologist independently delineated the ROIs for the same cases following the same protocol. Consistency between masks was quantitatively assessed using the Dice similarity coefficient (DSC), Jaccard similarity coefficient (JSC), and Hausdorff distance (HD) (22).

Feature extraction

In the initial approach utilizing 3D ROIs, resampling was performed to achieve a voxel size of 1 mm × 1 mm × 1 mm, and the grayscale was discretized into 25 levels directly on the entire ROI image. In the second method, 2D ROIs were generated automatically from the entire volume of the tumor area by selecting the slice with the largest cross-sectional area (23). The 2D ROIs were subsequently resampled to achieve a pixel size of 1 mm × 1 mm, with the image grayscale discretized into 25 levels. Radiomics features were extracted from both 2D ROIs and 3D ROIs. Overall, 1,032 and 1,095 radiomics features were extracted from 2D and 3D images, respectively. Further details on all features are available in Appendix 2. The feature calculation formula adheres to the Imaging Biomarker Standardization Initiative standard (24,25). It is based on the documentation of the Py_Radiomics software package (https://pyradiomics.readthedocs.io/en/latest/features.html). All the extracted radiomics features were continuous variables. The workflow of the radiomics analyses is shown in Figure 2.

Figure 2 Workflow of radiomics analysis and model evaluation. (A) ROI segmentation. Manual segmentation of the primary tumor on arterial-phase CECT in 2D (single axial slice with the largest cross-section) and 3D (whole-tumor volume). (B) Feature extraction. Radiomics features computed from the segmented ROI, including shape, first-order intensity statistics, and texture families (GLDM, GLCM, GLSZM, GLRLM, NGTDM). (C) Feature selection and model construction. Pipeline includes consistency analysis (reproducibility), Pearson correlation filtering, and univariate screening, followed by multivariable logistic regression to derive the radscore. (D) Model evaluation and survival analysis. Performance is illustrated by ROC curves (example AUC displayed) and Kaplan-Meier curves for RFS stratified by the Rad_2D score (NE-high vs. non-NE). 2D, two-dimensional; 3D, three-dimensional; AUC, area under the ROC curve; CECT, contrast-enhanced computed tomography; GLCM, Gray Level Co-occurrence Matrix; GLDM, Gray Level Dependence Matrix; GLRLM, Gray Level Run Length Matrix; GLSZM, Gray Level Size Zone Matrix; NE, neuroendocrine; NGTDM, Neighboring Gray Tone Difference Matrix; Rad_score, radiomics score; ROC, receiver operating characteristic; ROI, region of interest.

Feature selection and model construction

Univariate analysis was performed, and features with a significance level of P<0.05 remained in the model. Additionally, features with a correlation coefficient exceeding 0.9 were removed from the model. More details about selecting for both 2D and 3D radiomics analysis are provided in Appendix 3. The predictive models (Rad_2D and Rad_3D) were generated via the multivariate logistic regression algorithm.

To assess the independence of the radiomics signature, we additionally fitted a clinical logistic model (Clinical) and a combined logistic model (Combined_2D) based (Rad_2D + clinical). Clinical covariates for the baseline model were pre-specified as lymph nodal metastasis (LNM) (clinical N0/cN+) and clinical T (cT) stage, together with age, sex, smoking history, and maximal tumor diameter to represent general demographics/exposure and tumor burden. Other conventional CT signs (e.g., lobulation, spiculated, pleural retraction) were not included in the clinical-only baseline to limit overfitting under events-per-variable (EPV) constraints and to avoid collinearity with radiomics features that already encode morphological/texture information.

NE subtype assessment

Indicator selection

SCLC molecular typing markers (SCLC-A/N/P/Y). SCLC-A (1:200, ab74065, Abcam), SCLC-N (1:150, ab60704, Abcam), SCLC-P (1:300, bs-21046R, Bioss), and SCLC-Y (1:80, ab52771, Abcam) were used. The selected formalin-fixed, paraffin-embedded tissues were cut into 4 µm slices, and all the staining steps were performed via Roche automated immunohistochemical instruments (Roche Diagnostic Products Co., Ltd., Shanghai, China) according to recommended standard protocols. The corresponding tissues were selected for pre-experiment according to the manufacturer’s instructions. After passing quality control, the formal experiment was carried out.

Scoring criteria

Immunohistochemistry (IHC) staining results were evaluated by resident physicians with more than 3 years of experience in clinicopathological diagnosis. These evaluations were subsequently verified by senior professors. In the IHC staining indicators, the antibodies used to localize to the nucleus were SCLC-A and SCLC-P; the antibodies used to localize to the nucleus and cytoplasm included SCLC-N and SCLC-Y. The semiquantitative score of protein expression was calculated by multiplying the percentage of positive cells (%) and the staining intensity (0–3 grade), that is, the H score = (0–3) × (0–100%) × 100, and the range was 0–300 (26,27). Molecular subtyping was defined according to the highest H score of SCLC-A/N/P/Y protein expression (7,28,29). Finally, SCLC-A and SCLC-N patients are categorized into the NE group, whereas SCLC-Y and SCLC-P patients are classified into the non-NE group (4,6). The representative imaging examples are shown in Figure 3.

Figure 3 Representative arterial-phase CECT and immunohistochemistry illustrating molecular subtypes. (A,C) 65-year-old man, SCLC-A (ASCL1-dominant/NE-high). (A) Axial arterial-phase CECT shows a lobulated right lower-lobe mass (blue arrow). (C) IHC, ×200 shows diffuse staining pattern consistent with ASCL1-dominant/NE-high subtype (markers detailed in “Methods”). (B,D) 55-year-old woman, SCLC-P (POU2F3-dominant/non-NE). (B) Axial arterial-phase CECT shows a right lower-lobe mass (blue arrow). (D) IHC, ×200 shows staining pattern consistent with POU2F3-dominant/non-NE subtype (markers detailed in “Methods”). CECT, contrast-enhanced computed tomography; IHC, immunohistochemistry; NE, neuroendocrine; SCLC-A, ASCL1-dominant (NE-high) SCLC; SCLC-P, POU2F3-dominant (non-NE) SCLC.

Statistical analysis

Continuous variables were expressed as means ± standard deviation or medians [with range or interquartile range (IQR)] as appropriate and compared between groups using an independent-sample t-test or Mann-Whitney U test. Categorical variables were summarized as frequencies (percentages) and compared using the χ² test or Fisher’s exact test. Selection of variables for multivariable models followed the EPV principle (a limited number of covariates relative to the number of outcome events). Model performance was assessed with receiver operating characteristic (ROC) curve analysis; we calculated the area under the curve (AUC) for each model and, using the Youden index, determined the optimal probability cutoff to calculate the corresponding sensitivity, specificity, and accuracy. The DeLong test was used to compare AUCs between models. We also fit a multivariable logistic regression model combining the Rad_2D score with the clinical covariates defined above to evaluate the added value of radiomics. For the exploratory recurrence-free survival (RFS) analysis, a Cox proportional hazards model was constructed including the standardized Rad_2D score and LNM status (cN0 vs. cN+) as covariates. Hazard ratios (HRs) with 95% confidence intervals (CIs) were reported (for continuous variables, HRs reflect the risk per 1 standard deviation increase in the covariate). The proportional hazards assumption was verified using Schoenfeld residual tests. Where appropriate, performance metrics such as AUC and concordance index (C-index) were validated with bootstrap resampling to obtain 95% CIs. All statistical analyses were performed using R software (version 4.1.0) and Python (version 3.8.0). A two-tailed P value <0.05 was considered statistically significant.

Results

Patient characteristics

A total of 51 patients were included (male 60.8%, 31/51; female 39.2%, 20/51), with a median age of 56 years (range 30–76 years). Smoking history comprised never 41.2% (21/51) and current/former 58.8% (30/51). Molecular subtypes were SCLC-A 12 (23.5%), SCLC-N 27 (52.9%), SCLC-P 7 (13.7%), and SCLC-Y 5 (9.8%), which we aggregated as NE 39/51 (76.5%) and non-NE 12/51 (23.5%). Detailed clinical and pathological characteristics are summarized in Table 1.

Survival information

Among all patients, the median follow-up was 96 months (IQR: 23–120 months; range 7–225 months). A total of 29 individuals experienced endpoint events by the follow-up cutoff time, resulting in a recurrence or metastasis rate of 54.9% within 5 years. The distribution of RFS times and metastatic sites is provided in Appendix 4.

Conventional CT features

Baseline demographics and conventional CT signs were comparable between NE and non-NE groups (Table 1). In univariate analysis, no clinical variable nor conventional CT feature reached statistical significance for discriminating NE vs. non-NE in this cohort (Appendix 5). These results suggest that routine clinical descriptors and visual CT signs alone have limited stand-alone discriminative value, justifying the radiomics-based approach.

Inter-observer and intra-observer consistency

Across 30 cases, intra-observer reproducibility of semiautomatic ROI segmentation was excellent, with a median DSC of 0.994 (IQR 0.990–0.996; range 0.951–0.998) and median JSC of 0.987 (IQR 0.980–0.993; range 0.906–0.995); the median HD was 2.85 mm (IQR 2.25–4.54 mm; range 0.80–7.12 mm). Inter-observer agreement between two radiologists was satisfactory: median DSC 0.879 (IQR 0.795–0.918; range 0.596–0.965), median JSC 0.784 (IQR 0.660–0.849; range 0.425–0.932), and median HD 10.15 mm (IQR 5.62–14.01 mm; range 2.55–23.38 mm). Detailed results of accurate parameters based on semiautomatic segmentation are presented in Appendix 6.

Feature selection and model performance

The details and distribution of the selected features in the 2D and 3D analyses between the NE and non-NE groups are shown in Appendix 7 and Figure S1. In the context of 2D analysis, we identified three independent predictors: gray-level run length matrix RunEntropy (Glrlm_RunEntropy), first-order skewness, and gray-level dependence matrix LDHGLE (Gldm_LDHGLE). Moreover, in the 3D analysis, two independent predictors were identified, namely, Glrlm_RunEntropy and Gray level co-occurrence matrix ClusterShade (Glcm_ClusterShade). The performance of the remaining radiomics features is shown in Table 2. The results of the ROC curve analysis for discrimination of NEs vs. non-NEs with each selected radiomics feature are shown in Figure S2. Additionally, using this common feature Glrlm_RunEntropy, we validated the four classifications, yielding results with no significant differences (Figure S3).

The Rad_2D model (built from the three 2D features) achieved an AUC of 0.806 (95% CI: 0.666–0.945) with an overall accuracy of 68.6%, sensitivity of 64.1%, and specificity of 83.3% for predicting NE subtype. The Rad_3D model (using the two 3D features) showed a similar AUC of 0.784 (95% CI: 0.634–0.934), with accuracy 66.7%, sensitivity 61.5%, and specificity 83.3% (Table 3, Figure 4). The difference in AUC between the 2D and 3D radiomics models was not statistically significant (DeLong test, P=0.75). In an exploratory survival analysis stratified by the radiomics predictions, the Rad_2D model demonstrated prognostic relevance: patients classified as “NE subtype” by Rad_2D had a 5-year RFS of 48.1%, compared to 62.5% in those predicted as “non-NE subtype” (log-rank P=0.03, Figure 5). In contrast, the Rad_3D-based stratification showed no significant RFS difference between its predicted NE vs. non-NE groups (5-year RFS 53.8% vs. 56.0%, log-rank P=0.31).

Figure 4 Comparison of ROC curves for Rad_2D and Rad_3D radiomics models. ROC curves comparing the diagnostic performance of the 2D radiomics model (Rad_2D, orange line) and the 3D radiomics model (Rad_3D, green line). Reported AUC (95% CI): Rad_2D =0.806 (0.666–0.945); Rad_3D =0.784 (0.634–0.934). AUC, area under the ROC curve; CI, confidence interval; Rad_2D, two-dimensional radiomics model; Rad_3D, three-dimensional radiomics model; ROC, receiver operating characteristic.

Figure 5 Kaplan-Meier survival by predicted molecular subtype using 2D vs. 3D radiomics scores. (A) Rad_2D score. Kaplan-Meier curves for RFS stratified by predicted subtype: pNEs (blue) vs. pnon-NEs (yellow). 5-year RFS: 48.1% vs. 62.5%; log-rank P=0.03. (B) Rad_3D score. Kaplan-Meier curves for RFS stratified by predicted subtype: pNEs (blue) vs. pnon-NEs (yellow). 5-year RFS: 53.8% vs. 56.0%; log-rank P=0.31. 2D, two-dimensional; 3D, three-dimensional; pNEs, predicted neuroendocrine-subtype group; pnon-NEs, predicted non-neuroendocrine-subtype group; Rad_score, radiomics score; RFS, recurrence-free survival.

To further validate the added value of the radiomics signature, we performed a post hoc comparison with clinical models. A logistic regression model based on clinical factors alone (LNM status, cT stage, age, sex, smoking history, and tumor maximum diameter) yielded an AUC of 0.673 (95% CI: 0.486–0.846) for NE subtype prediction. When the Rad_2D radiomics score was combined with these clinical variables in a single model, the Combined_2D model achieved an AUC of 0.861 (95% CI: 0.737–0.961) (Figure 4). In this combined model, the Rad_2D score remained an independent predictor of NE subtype with an adjusted odds ratio (OR) of 6.22 (95% CI: 1.73–22.39, P=0.005). The AUC of the combined model was higher than that of either the clinical model or the Rad_2D model alone, supporting the incremental predictive value of the radiomics signature over standard clinical features.

RFS (exploratory)

In the exploratory analysis of RFS, we used a Cox proportional hazards model including the Rad_2D score and LNM status. The Rad_2D score showed a trend toward shorter RFS: each one-standard deviation increase in the Rad_2D score corresponded to an estimated HR of 1.46 for recurrence (95% CI: 0.95–2.24, P=0.08). LNM status was not significantly associated with RFS in this model (HR for cN+ vs. cN0: 0.93, 95% CI: 0.42–2.05, P=0.85). The model’s ability to discriminate RFS was modest (C-index 0.596, bootstrap 95% CI: 0.332–0.671). There was no evidence of violation of the proportional hazards assumption (based on Schoenfeld tests).

Discussion

In this preliminary study, we demonstrated that quantitative radiomics analysis of CECT images can noninvasively differentiate the NE molecular subtype of SCLC from the non-NE subtype. To our knowledge, this is the first study to use quantitative radiomics features to assess the molecular subtypes of SCLC. The key finding is that a radiomics signature derived from CECT of the primary tumor correlates strongly with the tumor’s molecular phenotype. Our 2D radiomics model—built on a single representative axial slice—achieved an AUC of 0.81 for distinguishing NE vs. non-NE subtypes, which was comparable to the performance of the 3D radiomics model (AUC of 0.78), with the Rad_2D model additionally providing prognostic value for RFS (P=0.03). These results suggest that subtle imaging patterns captured by radiomics features reflect underlying molecular distinctions that are imperceptible to human readers, underscoring the potential of radiomics as a noninvasive plan for SCLC.

Our findings align with the growing body of radio-genomic evidence indicating that medical images encode tumor biology. Prior studies in lung adenocarcinoma have shown that CT radiomics features can noninvasively predict specific driver mutations (e.g., EGFR, ALK) and histologic subtypes. However, in SCLC, such radiomics applications have been scarce. Previous radiomics efforts in SCLC have focused mainly on classifying broad histologic categories (30). Our study fills this gap by demonstrating that radiomics can capture the intra-SCLC heterogeneity associated with NE differentiation. This is clinically important because NE-high SCLC (driven by SCLC-A/N expression) is known to exhibit more aggressive behavior and poorer prognosis than non-NE subtypes (SCLC-Y/P) (6). Additionally, our results showed that clinical factors and conventional CT descriptors failed to significantly distinguish between NE and non-NE groups, underscoring the limitations of visual interpretation. In contrast, radiomics analysis provides a reproducible, noninvasive method that can infer molecular subtypes directly from imaging, potentially enhancing early prognostic stratification and informing personalized treatment strategies in SCLC.

It is noteworthy that the radiomics features selected in our models hint at biological differences between NE and non-NE SCLC. In both 2D and 3D analyses, Glrlm_RunEntropy emerged as a top predictor, suggesting that NE tumors have more complex, heterogeneous texture patterns on CT. Higher run-length entropy indicates greater randomness in the distribution of continuous pixel intensity runs, which could reflect the chaotic architecture and necrosis often seen in high-grade NE tumors. Similarly, Firstorder_Skewness (selected in the Rad_2D model) captures asymmetry in the intensity histogram (31); a skewed distribution might correspond to irregular tumor attenuation present in certain subtypes. Another 2D feature, Gldm_LDHGLE (large dependence high gray-level emphasis), points to differences in regional intensity uniformity (32)—potentially indicating that NE tumors contain areas of densely packed cells (high gray-level) with large zones of similar intensity. While the precise histopathologic correlates of these metrics remain to be validated, our results support the concept that radiomics can reveal phenotypic nuances beyond the scope of conventional CT criteria.

A surprising but practical result of this study is that 2D radiomics analysis performed on par with 3D analysis for subtype classification. One might expect that 3D features (analyzing the entire tumor volume) would carry more information and thus outperform 2D (single-slice) features. However, our data showed no significant difference in AUC between the Rad_2D and Rad_3D models (P=0.75). This finding is consistent with several recent reports (33-38) and 2D radiomics models considering the time and manpower required for 3D segmentation are recommended for clinical applications because of their practicality and efficiency (33). Our results echo these observations—the single largest cross-sectional slice often contains the most informative tumor morphology and texture (e.g., the maximal heterogeneity), whereas adding many peripheral slices in 3D may introduce noise from partial-volume effects or less cellular tumor regions. The practical implication is that a 2D radiomics approach might suffice for this task, enabling faster analysis and easier standardization across institutions (since manual 3D segmentation is more time-consuming and subject to variability). This is encouraging for clinical translation, as 2D radiomics can be more readily integrated into workflows without requiring extensive volumetric processing.

From a clinical perspective, a noninvasive imaging-based tool for SCLC subtyping could offer substantial value. Currently, the identification of NE subtypes relies on immunohistochemical analysis of invasive biopsy or surgical specimens—procedures that carry inherent risks such as pneumothorax or hemorrhage, and may be impractical for patients with advanced disease or poor performance status. If externally validated, our radiomics model could provide a pre-therapeutic estimation of molecular subtype, thereby guiding treatment decisions when tissue sampling is infeasible. Moreover, even when biopsy is feasible, limited sampling may fail to capture the intratumoral heterogeneity. In contrast, radiomics evaluates the entire tumor (or its most representative region), potentially offering a more comprehensive view of the tumor’s underlying biology.

Beyond subtype classification, accurate prognostic assessment is also essential for implementing personalized treatment strategies, particularly given the considerable variability in clinical outcomes among SCLC patients—even within the same clinical stage and treatment regimen (39). Our exploratory findings suggest that the radiomics signature may hold prognostic value. Prior studies have demonstrated the prognostic utility of radiomics features in non-SCLC (40,41), as well as in predicting RFS in SCLC patients after chemotherapy (39,42). In our cohort, the Rad_2D score stratified patients into NE and non-NE groups with significantly different RFS. NE subtypes are known to exhibit higher proliferative activity (Ki-67 index) and greater resistance to conventional therapies (7), providing a biological rationale for their shorter RFS. Although the Rad_2D score showed only marginal significance in multivariable Cox analysis, this trend remains consistent with the known aggressiveness of NE subtypes. Notably, the 2D model focuses on the largest cross-sectional tumor area, which often contains the most aggressive subregions (e.g., necrotic cores, vascular invasion zones), driving recurrence. This finding aligns with reports in NSCLC, where 2D radiomics often outperforms 3D analysis in survival prediction due to reduced noise and artifact sensitivity (34). Taken together, the significant RFS divergence observed in our study supports the potential of 2D radiomics as a practical, noninvasive prognostic tool in SCLC.

There are several limitations in this study. First, it was conducted retrospectively at a single center with a relatively small sample size, which may limit the generalizability of the findings. Second, although the radiomics models demonstrated encouraging performance, they lack external and prospective validation, which is necessary before clinical implementation. Third, the analysis was based solely on CT radiomics, without integrating other potentially informative modalities such as positron emission tomography (PET) imaging or genomic data. Future studies with larger, multicenter cohorts and multimodal approaches are needed to confirm and extend these preliminary findings.

Conclusions

In summary, this study demonstrates the potential of CT-based radiomics as a noninvasive tool for subtyping SCLC, effectively distinguishing NE from non-NE molecular subtypes. Notably, 2D radiomics performed comparably to 3D while additionally showing promise in stratifying RFS, offering a more practical and clinically accessible approach for future implementation.

Acknowledgments

We thank Jian-Wei Li for his help in completing this manuscript.

Footnote

Reporting Checklist: The authors have completed the STARD reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1041/rc

Data Sharing Statement: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1041/dss

Peer Review File: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1041/prf

Funding: This research was funded by the Scientific Research Project of Shanxi Provincial Health Commission (No. 2024099), National High Level Hospital Clinical Research Funding (No. 80102022505), the National Natural Science Foundation of China (No. 81601494) and Beijing Hope Run Special Fund of Cancer Foundation of China (No. LC2022A22).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1041/coif). J.L.R. is an employee of GE HealthCare, Beijing, China. The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was approved by Ethics Committee of the Cancer Hospital, Chinese Academy of Medical Sciences (approval No. NCC2016G-029). Informed consent was waived due to the retrospective nature. All procedures were conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
2. Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 2021;71:209-49. [Crossref] [PubMed]
3. Kim SY, Park HS, Chiang AC. Small Cell Lung Cancer: A Review. JAMA 2025;333:1906-17. [Crossref] [PubMed]
4. Megyesfalvi Z, Gay CM, Popper H, et al. Clinical insights into small cell lung cancer: Tumor heterogeneity, diagnosis, therapy, and future directions. CA Cancer J Clin 2023;73:620-52. [Crossref] [PubMed]
5. Thai AA, Solomon BJ, Sequist LV, et al. Lung cancer. Lancet 2021;398:535-54. [Crossref] [PubMed]
6. Rudin CM, Poirier JT, Byers LA, et al. Molecular subtypes of small cell lung cancer: a synthesis of human and mouse model data. Nat Rev Cancer 2019;19:289-97. [Crossref] [PubMed]
7. Baine MK, Hsieh MS, Lai WV, et al. SCLC Subtypes Defined by ASCL1, NEUROD1, POU2F3, and YAP1: A Comprehensive Immunohistochemical and Histopathologic Characterization. J Thorac Oncol 2020;15:1823-35. [Crossref] [PubMed]
8. Rudin CM, Brambilla E, Faivre-Finn C, et al. Small-cell lung cancer. Nat Rev Dis Primers 2021;7:3. [Crossref] [PubMed]
9. Redin E, Quintanal-Villalonga Á, Rudin CM. Small cell lung cancer profiling: an updated synthesis of subtypes, vulnerabilities, and plasticity. Trends Cancer 2024;10:935-46. [Crossref] [PubMed]
10. Megyesfalvi Z, Barany N, Lantos A, et al. Expression patterns and prognostic relevance of subtype-specific transcription factors in surgically resected small-cell lung cancer: an international multicenter study. J Pathol 2022;257:674-86. [Crossref] [PubMed]
11. Lee MV, Katabathina VS, Bowerson ML, et al. BRCA-associated Cancers: Role of Imaging in Screening, Diagnosis, and Management. Radiographics 2017;37:1005-23. [Crossref] [PubMed]
12. Cox VL, Bhosale P, Varadhachary GR, et al. Cancer Genomics and Important Oncologic Mutations: A Contemporary Guide for Body Imagers. Radiology 2017;283:314-40. [Crossref] [PubMed]
13. Cozzi D, Bicci E, Cavigli E, et al. Radiomics in pulmonary neuroendocrine tumours (NETs). Radiol Med 2022;127:609-15. [Crossref] [PubMed]
14. Bicci E, Cozzi D, Cavigli E, et al. Reproducibility of CT radiomic features in lung neuroendocrine tumours (NETs) patients: analysis in a heterogeneous population. Radiol Med 2023;128:203-11. [Crossref] [PubMed]
15. Ogut E. Artificial Intelligence in Clinical Medicine: Challenges Across Diagnostic Imaging, Clinical Decision Support, Surgery, Pathology, and Drug Discovery. Clin Pract 2025;15:169. [Crossref] [PubMed]
16. Lambin P, Woodruff HC, Mali SA, et al. Radiomics Quality Score 2.0: towards radiomics readiness levels and clinical translation for personalized medicine. Nat Rev Clin Oncol 2025;22:831-46. [Crossref] [PubMed]
17. Warkentin MT, Al-Sawaihey H, Lam S, et al. Radiomics analysis to predict pulmonary nodule malignancy using machine learning approaches. Thorax 2024;79:307-15. [Crossref] [PubMed]
18. Felfli M, Liu Y, Zerka F, et al. Systematic Review, Meta-Analysis and Radiomics Quality Score Assessment of CT Radiomics-Based Models Predicting Tumor EGFR Mutation Status in Patients with Non-Small-Cell Lung Cancer. Int J Mol Sci 2023;24:11433. [Crossref] [PubMed]
19. Ottaiano A, Grassi F, Sirica R, et al. Associations between Radiomics and Genomics in Non-Small Cell Lung Cancer Utilizing Computed Tomography and Next-Generation Sequencing: An Exploratory Study. Genes (Basel) 2024;15:803. [Crossref] [PubMed]
20. Chen W, Lin G, Feng Y, et al. Intratumoral and peritumoral CT radiomics in predicting anaplastic lymphoma kinase mutations and survival in patients with lung adenocarcinoma: a multicenter study. Cancer Imaging 2025;25:35. [Crossref] [PubMed]
21. Demircioğlu A. The effect of feature normalization methods in radiomics. Insights Imaging 2024;15:2. [Crossref] [PubMed]
22. Liu Z, Mhlanga JC, Xia H, et al. Need for Objective Task-Based Evaluation of Image Segmentation Algorithms for Quantitative PET: A Study with ACRIN 6668/RTOG 0235 Multicenter Clinical Trial Data. J Nucl Med 2024;65:485-92. [Crossref] [PubMed]
23. Zhang X, Zhang G, Qiu X, et al. Radiomics under 2D regions, 3D regions, and peritumoral regions reveal tumor heterogeneity in non-small cell lung cancer: a multicenter study. Radiol Med 2023;128:1079-92. [Crossref] [PubMed]
24. Whybra P, Zwanenburg A, Andrearczyk V, et al. The Image Biomarker Standardization Initiative: Standardized Convolutional Filters for Reproducible Radiomics and Enhanced Clinical Insights. Radiology 2024;310:e231319. [Crossref] [PubMed]
25. Zwanenburg A, Vallières M, Abdalah MA, et al. The Image Biomarker Standardization Initiative: Standardized Quantitative Radiomics for High-Throughput Image-based Phenotyping. Radiology 2020;295:328-38. [Crossref] [PubMed]
26. McCarty KS Jr, Miller LS, Cox EB, et al. Estrogen receptor analyses. Correlation of biochemical and immunohistochemical methods using monoclonal antireceptor antibodies. Arch Pathol Lab Med 1985;109:716-21.
27. Specht E, Kaemmerer D, Sänger J, et al. Comparison of immunoreactive score, HER2/neu score and H score for the immunohistochemical evaluation of somatostatin receptors in bronchopulmonary neuroendocrine neoplasms. Histopathology 2015;67:368-77. [Crossref] [PubMed]
28. Qi J, Zhang J, Liu N, et al. Prognostic Implications of Molecular Subtypes in Primary Small Cell Lung Cancer and Their Correlation With Cancer Immunity. Front Oncol 2022;12:779276. [Crossref] [PubMed]
29. Qu S, Fetsch P, Thomas A, et al. Molecular Subtypes of Primary SCLC Tumors and Their Associations With Neuroendocrine and Therapeutic Markers. J Thorac Oncol 2022;17:141-53. [Crossref] [PubMed]
30. Guo Y, Song Q, Jiang M, et al. Histological Subtypes Classification of Lung Cancers on CT Images Using 3D Deep Learning and Radiomics. Acad Radiol 2021;28:e258-66. [Crossref] [PubMed]
31. van Griethuysen JJM, Fedorov A, Parmar C, et al. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res 2017;77:e104-7. [Crossref] [PubMed]
32. Bashir U, Siddique MM, Mclean E, et al. Imaging Heterogeneity in Lung Cancer: Techniques, Applications, and Challenges. AJR Am J Roentgenol 2016;207:534-43. [Crossref] [PubMed]
33. Li Y, Yang L, Gu X, et al. Computed tomography radiomics identification of T1-2 and T3-4 stages of esophageal squamous cell carcinoma: two-dimensional or three-dimensional? Abdom Radiol (NY) 2024;49:288-300. [Crossref] [PubMed]
34. Yang G, Nie P, Zhao L, et al. 2D and 3D texture analysis to predict lymphovascular invasion in lung adenocarcinoma. Eur J Radiol 2020;129:109111. [Crossref] [PubMed]
35. Meng L, Dong D, Chen X, et al. 2D and 3D CT Radiomic Features Performance Comparison in Characterization of Gastric Cancer: A Multi-Center Study. IEEE J Biomed Health Inform 2021;25:755-63. [Crossref] [PubMed]
36. Shen C, Liu Z, Guan M, et al. 2D and 3D CT Radiomics Features Prognostic Performance Comparison in Non-Small Cell Lung Cancer. Transl Oncol 2017;10:886-94. [Crossref] [PubMed]
37. Ahn SJ, Kim JH, Park SJ, et al. Prediction of the therapeutic response after FOLFOX and FOLFIRI treatment for patients with liver metastasis from colorectal cancer using computerized CT texture analysis. Eur J Radiol 2016;85:1867-74. [Crossref] [PubMed]
38. Lubner MG, Stabo N, Lubner SJ, et al. CT textural analysis of hepatic metastatic colorectal cancer: pre-treatment tumor heterogeneity correlates with pathology and clinical outcomes. Abdom Imaging 2015;40:2331-7. [Crossref] [PubMed]
39. Zheng X, Liu K, Li C, et al. A CT-based radiomics nomogram for predicting the progression-free survival in small cell lung cancer: a multicenter cohort study. Radiol Med 2023;128:1386-97. [Crossref] [PubMed]
40. Huang Y, Liu Z, He L, et al. Radiomics Signature: A Potential Biomarker for the Prediction of Disease-Free Survival in Early-Stage (I or II) Non-Small Cell Lung Cancer. Radiology 2016;281:947-57. [Crossref] [PubMed]
41. Wang T, She Y, Yang Y, et al. Radiomics for Survival Risk Stratification of Clinical and Pathologic Stage IA Pure-Solid Non-Small Cell Lung Cancer. Radiology 2022;302:425-34. [Crossref] [PubMed]
42. Huang WP, Liu SY, Han YJ, et al. Development of CT-Based Imaging Signature for Preoperative Prediction of Invasive Behavior in Pancreatic Solid Pseudopapillary Neoplasm. Front Oncol 2021;11:677814. [Crossref] [PubMed]